跳转到内容

Perceptual Quality

perceptual_quality 衡量预测结果自身的无参考视觉质量,不将 prediction 与 GT reference 做逐帧配对比较。它关注锚点帧层面的锐度、干净程度与审美合理性,不评估语义正确性、prompt 对齐或几何/物理合规性。

WorldAtlas 将两类互补的无参考 IQA 信号融合为单一感知质量分数:CLIP-IQA+ 刻画整体视觉自然度与感知协调性;LAION 美学预测器刻画审美吸引力。二者均在 prediction anchor 帧上独立打分后再聚合。

CLIP-IQA+ 基于 CLIP 视觉—语言表征评估单帧图像的整体感知质量。对锚点帧 IiI_i,模型输出 native 分数 qinativeq_i^{\mathrm{native}},归一化为:

siiqa=clip ⁣(qinative1, 0, 1)(1)s_i^{\mathrm{iqa}} = \mathrm{clip}\!\left(\frac{q_i^{\mathrm{native}}}{1},\, 0,\, 1\right) \tag{1}

美学分数沿用 CLIP ViT-L/14 特征与 LAION-Aesthetics 线性回归头(pyiqa laion_aes)。对锚点帧 IiI_i:

siaes=clip ⁣(g(norm(CLIP(Ii)))10, 0, 1)(2)s_i^{\mathrm{aes}} = \mathrm{clip}\!\left(\frac{g(\mathrm{norm}(\mathrm{CLIP}(I_i)))}{10},\, 0,\, 1\right) \tag{2}

每帧感知质量取两组件的算术平均:

si=siiqa+siaes2(3)s_i = \frac{s_i^{\mathrm{iqa}} + s_i^{\mathrm{aes}}}{2} \tag{3}

设协议选出 NN 个 prediction anchor 帧,指标 raw 值为:

Spq=1N∑i=1Nsi(4)S_{\mathrm{pq}} = \frac{1}{N}\sum_{i=1}^{N} s_i \tag{4}

Audit manifest: docs/assets/metric_examples/curated_same_case_manifest.json.

输入图像: image_static_photorealistic_aquatic_landscape_0079
输入图像
输入已核验 · 同一 benchmark case · gap 0.3998

image_static_photorealistic_aquatic_landscape_0079

套件: image_static场景: photorealistic / outdoor / aquatic_landscape
基准 prompt target

Camera move right. A stone archway with pathway, grassy area, and mountains.

prompt source: prompt_target

两条视频的首帧都已通过与展示输入的几何匹配核验;高低标签只比较当前 metric。

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较高指标分分数 0.7155

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较低指标分分数 0.3157
输入图像: image_static_stylized_aquatic_landscape_0110
输入图像
输入已核验 · 同一 benchmark case · gap 0.3910

image_static_stylized_aquatic_landscape_0110

套件: image_static场景: stylized / outdoor / aquatic_landscape
基准 prompt target

Camera orbit right. An aquatic landscape with oil rig, industrial buildings, and rocky platform.

prompt source: prompt_target

两条视频的首帧都已通过与展示输入的几何匹配核验;高低标签只比较当前 metric。

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较高指标分分数 0.6485

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

较低指标分分数 0.2576

BibTeX — CLIP-IQA+
@inproceedings{wang2022exploring,
author = {Wang, Jianyi and Chan, Kelvin CK and Loy, Chen Change},
title = {Exploring CLIP for Assessing the Look and Feel of Images},
booktitle = {AAAI},
year = {2023}
}